[Ffmpeg-devel] Bug in snow revision 1.50 - broken lossless mode

[Yartrebo]

On Wed, 2005-04-13 at 20:52 +0200, Michael Niedermayer wrote:
> Hi
> > In that case, it would be nice to make 5/3 the default for lossless. It
> > really doesn't make snow look good to have the default mode (9/7,
> > lossless) look so ugly.
> 
> the default of option A should not depend on option B, its confusing and 
> problematic, there should be a check for this and the encoder should die with 
> an error if 9/7 is used in lossless mode (patch welcome)
> 
> [...]
I've been testing reducing the levels of decomposition for lossless mode
(5/3 wavelet) (eliminating dwt/idwt completely would require a little
more hacking) and I've got some interesting results:

temp_speak_up is a 512x384 encode (cartoon).
temp is a 512x384 transcode of a high quality xvid (CG animation).
temp_gummi is a 320x240 transcode of a low quality divx (cartoon).
All are 3 seconds long.

temp1 has 5 decomposition levels.
temp2 has 4 decomposition levels.
...
temp5 has 1 decomposition level.

size    name
5808338 temp_speak_up_1.avi
5806786 temp_speak_up_2.avi
5802532 temp_speak_up_3.avi
5802568 temp_speak_up_4.avi
5800562 temp_speak_up_5.avi

3567262 temp1.avip1.avi
3565508 temp2.avi
3563214 temp3.avi
3557608 temp4.avi
3532662 temp5.avi

2619304 temp_gummi_1.avi
2619496 temp_gummi_2.avi
2620204 temp_gummi_3.avi
2617298 temp_gummi_4.avi
2553554 temp_gummi_5.avi

This indicates that lossless compression is better without dwt/idwt than
with it. It would also make lossless compression substantially faster to
remove dwt/idwt.

So I suggest not doing dwt/idwt for lossless mode. We'd even get two
birds with one stone, since the rescaled coefficients wouldn't be
causing any more troubles. Any comments or ideas about this?

Robert